A neural based system for obstacle detection and scene reconstruction

dA stereo vision based obstacle detection system is presented. The matching process on the input stereogram is performed as an optimisation of an energy functional through a variational approach yielding dense disparity maps. The energy minimisation is implemented by a Cellular Neural Network. The state of the art of the hardware implementation of the system is presented. Some experiments on the use of the system in outdoors applications are shown. These tests demonstrate the feasibility of an obstacle detection system for an autonomous surveillance robotic platform. The real time characteristics of the hardwired version of the algorithm will allow the temporal, and spatial, integration of data, with a considerable reduction in the otherwise unavoidable data noise.